An Improved Latent Fingerprint's Feature Extraction And Matching Algorithm

Abstract

Latent fingerprints are the impressions of partial ridges left on the surface of objects touched unintentionally at crime scenes and constitute a valuable source of evidence in law enforcement agencies to help solve crimes. However, majority of the processing (marking region of interest (ROI), singular points (SP), orientation field and minutiae points) for latent prints identification is done manually by forensic experts. The existing methods involve forensic experts to manually mark the features in latent and then input it to the system for automatic matching with reference prints. The matcher returns a list of candidates that are manually checked by experts to take final decision. The practice of manually marking features in latents is laborious, time consuming and human dependent which may results in wrong identification. Therefore there is a need to automate this process to avoid aforesaid constraints.

This thesis describes an automated approach of segmentation and enhancement for latent fingerprints identification. Currently, a few attempts have been made in this respect and still remain a challenging problem due to: (i) poor quality, (ii) small friction ridge area, (iii) presence of non-linear distortion, (iv) blurring or smudging, and (v) complex background noise.

In this research, an algorithm for automated segmentation of latent fingerprints is proposed. The latent image data is classified into clusters using K-means clustering technique which results in pixels having similar characteristics to fit in one cluster (foreground) while pixels having opposite characteristics to other cluster (background). Tophat filtering is applied to enhance the clustered data and mask is generated on the basis of this enhanced information. Segmentation is achieved by applying the generated mask on latent image. The proposed algorithm for segmentation of latent fingerprints is automated without any sort of human involvement. Performance of proposed algorithm is evaluated by computing the missed detection rate (MDR) and false detection rate (FDR) and comparison of proposed method with other existing algorithms is done. Simulation results on NIST SD-27 (database of latent fingerprint images containing 258 latent fingerprints along with their mated rolled prints) show significant performance enhancement of proposed method having average MDR and FDR of 4.77% and 26.06% respectively. Furthermore, subjective comparison is made using visual segmentation reliability (VSR) which is the ratio of intersectional area of automated and ground truth latent to manually marked segmented latent. VSR approaches to 90% for good quality images, 70-80% for bad quality images and 50-60% for ugly quality images. Matching performance is improved when the segmented input is applied to commercial-off-the-shelf (COTS tenprint) matcher as compared with un- segmented input.

Another contribution of proposed research is towards the enhancement of latent fingerprints. Enhancement of segmented latent is performed using Gabor filter bank. It has five image-dependent-parameters like orientation ∅ , standard deviations 𝜎� and 𝜎� of the Gaussian function, time period T and the convolution mask size. The selection of these parameters plays a crucial role in fingerprint enhancement specifically the orientation ∅

and standard deviation 𝜎� and 𝜎� . The latent image is divided into blocks of W x W centered at pixel (i, j) and gradients ��� and ��� along x-axis and y-axis are computed by applying Sobel operator at every pixel. Orientation ∅ is computed on the basis of these computed gradients. Ridge frequency F(i, j) is estimated by calculating the grey level value of each pixel, housed in the block, and is projected in a direction perpendicular to the local ridge orientation and ridge spacing S(i, j). An improvement in frequency estimation is achieved by introducing Gaussian low pass filter that minimizes the noise levels. Ridge orientation ∅ and frequency F(i, j) is used to design an even-symmetric Gabor filter. Spatial convolution of the latent fingerprint with Gabor filter is performed to generate enhanced

latent image. Simulation results on NIST SD-27 show that improvement in matching is increased 11% in comparison to automated latent fingerprint segmentation and enhancement algorithm by Zhang et al in 2013.